Recursos de las OMAPED

Although a formal evaluation of the accuracy of variant calling pipelines remains unfeasible for nonsimulated sequence data (Li, 2014), we estimated the performance of the workflow using both sorghum and Arabidopsis sequence data. For the sorghum data, we compared the variants called from sorghum WGS data via the RIG workflow to (i) a collection of reliable variants that were not used to train the VQSR models and (ii) a previously published sorghum variant calling analysis. We then used the sorghum WGS variants to recalibrate reduced representation data, and used the re- calibrated variants for a genome-wide association study. Lastly, we further validated the performance of the RIG workflow using publicly available Sanger sequence and WGS data from Arabidopsis. Evaluation of recalibrated sorghum variants:

First, we examined the overlap between the recalibrated sorghum WGS variants and a collection of reliable variants that were not used to train the VQSR models. This collection of reliable variants, hereafter referred to as the Independent-Family (IF) set, originated from a biparental cross genotyped using a reduced representation method; the IF set was obtained in a similar manner to the Family Reference Variant Resource that was used for training during VQSR, and the IF set represented a set of highly specific, genetically mappable variants (see the section Materials and Methods). Of the 10,737 SNPs and 3740 indels in the IF set, 10,557 SNPs and 3632 indels had also been called from the 49 WGS samples (of which 2 samples represented the parents of the biparental cross). The IF variants present in the recalibrated

Figure 2.5: Interrelation of different genomic sequence data sources using the RIG workflow. (A) Schematic of how variants from reduced representation sequence (RR) data present in whole-genome sequence (WGS) data can be used to VQSR the WGS raw variants and assign VQSLOD scores to those variants. (BD) Visualization of the genomic region of Sb07g003860, a gene involved in sorghum midrib coloration (Bout and Vermerris, 2003). (B) the Sbi1.4 gene annotation; (C) shows the assigned VQSLOD scores for variants called in the region from WGS data; (D) shows the depth of coverage and mapped sequence reads for reduced representation and WGS data, respectively, for one sorghum line (BTx642). The RIG workflow enables vari- ants called in the reduced representation sequence data to be used to inform and recalibrate the WGS analyses, and vice versa. This puts all of the variant calls into the GATKs probabilistic framework whereby variants can be filtered based on their reliability. Users interested in more sensitive or specific call sets can choose more inclusive or exclusive tranches, respectively, by changing the cutoff indicated by the blue dotted line in Panel C. The common and standardized file formats emitted by the GATK enable downstream interoperability between analysis and visualization tools, such as the Integrative Genomics Viewer that produced (B) and (D) (Thor- valdsd´ottir et al., 2013). RIG, Recalibration and Interrelation of genomic sequence data with the GATK; VQSR, Variant Quality Score Recalibration; VQSLOD, loga- rithm of odds ratio that a variant is real vs. not under the trained Gaussian mixture model; GATK, Genome Analysis Toolkit.

WGS variants had median VQSLOD scores of 8.22 and 5.29 for SNPs and indels, respectively, suggesting that the trained Gaussian mixture models correctly assigned true variants with highly positive VQSLOD scores (Figure A.1, Table A.1, and Table A.2). Furthermore, the proportion of IF set variants that were also contained in the 95% and 75% tranches correspond to their respective tranche cutoffs, indicating that the tranche cutoffs were functioning as expected. Since tranche cutoffs represent the VQSLOD score over which a certain proportion of variants from the designated VQSR truth set will be retained, we expected the proportion of IF variants present in each tranche to approximate the tranche cutoff. As expected, proportions of the IF set retained in each tranche were similar to the tranche cutoff. For example, the 95% SNP tranche retained 97% of the SNPs in the IF set, and the 95% indel tranche retained 94% of the indels in the IF set (Table A.2). These results indicate that the Gaussian mixture models for the WGS data were adequately trained and that the tranche cutoffs were functioning as expected.

Second, we compared the recalibrated sorghum WGS variants to a previously published sorghum variant calling analysis. The previous study from Mace et al. (2013) called SNPs and indels from 47 sorghum WGS samples; the SNP calls were recently made available as part of Gramene build 42 (accessed September 2014), hereafter referred to as the Gramene42-Mace2013 set (Monaco et al., 2014). After excluding noncomparable variants from the calls produced by the RIG workflow (i.e., indels, SNPs on super contigs, and variants not found in the 47 samples), we obtained a Raw set comprised of 18,160,612 SNPs. We constructed an additional two sets from this Raw set for comparison: the Sensitive set, derived from the 95% tranche and comprised of 8,071,250 SNPs, and the Specific set, derived from the 75% tranche and comprised of 3,353,064 SNPs. Of the 6,450,628 SNPs in Gramene42- Mace2013 set, 5,002,099 were present in the Raw set. It is difficult to conclusively

attribute the 1,448,529 SNP difference to any specific factors, and high discordance between different variant callers is not uncommon (O’Rawe et al., 2013); we note that Mace et al. (2013) did not perform BQSR nor realignment around indels prior to calling SNPs, and they also used a different SNP calling algorithm. The overlapping 5,002,099 SNPs were used to compare the distribution of VQSLOD scores between the four sets (Figure 2.6). Because the VQSLOD score of all of the SNPs in the comparison were assigned under the same Gaussian mixture model and because the model was adequately trained as shown by the IF validation, comparisons of the relative sensitivity and specificity between the sets can be made. Given two sets of variants with similar VQSLOD distributions, the larger of the two sets contains more variants that are as likely to be true positives than the smaller set and is thus more sensitive. Furthermore, given two sets of variants where the VQSLOD distribution of one set contains a greater proportion of high VQSLOD score variants, the set with the greater proportion of high VQSLOD score variants contains variants that are more likely to be true positives and is thus more specific. As such, we find that the Raw set is the most sensitive but least specific; correspondingly, the Specific set is the most specific but least sensitive (Figure 2.6). The Sensitive set produced by the RIG workflow shows a dramatic improvement over the Gramene42-Mace2013 set in that it contains 1,620,622 more SNPs than the Gramene42-Mace2013 set while the median VQSLOD score remains similar with fewer negative VQSLOD scores, suggesting that the RIG workflow enabled greatly increased sensitivity without a corresponding loss in specificity.

As a final validation of the workflow with sorghum variants, we used a set of variants from reduced representation sequence data that had been recalibrated with WGS data to reproduce genome wide association results from the sorghum literature. There were 171 individuals contained within our reduced representation samples that

Figure 2.6: Comparison of VQSLOD score distributions for RIG-produced variant sets and a variant set from a previous study. VQSLOD (log of odds that a variant is real vs. not under the trained Gaussian mixture model) scores were calculated during VQSR of SNPs found in whole-genome sequence data using a Gaussian mixture model trained using SNPs originally found in reduced representation sequence data. For the 5,002,099 SNPs from Gramene42-Mace2013 that had been assigned VQSLOD scores in the Raw set produced by the RIG workflow, the median VQSLOD score is similar to the median of the 8,071,250 SNPs in the Sensitive set. The Sensitive set contains 1,620,622 more SNPs than the 6,450,628 SNPs in Gramene42-Mace2013, suggesting that the RIG-enabled VQSR allowed for a considerably more sensitive call set without a corresponding loss in specificity. VQSLOD, logarithm of odds ratio that a variant is real vs. not under the trained Gaussian mixture model; RIG, Recalibration and Interrelation of genomic sequence data with the GATK; VQSR, Variant Quality Score Recalibration; SNP, single-nucleotide polymorphism.

had also previously been phenotyped as part of a sorghum association panel (Brown et al., 2008). After recalibrating the reduced representation data with the WGS data, we used the genotypes for these 171 individuals and phenotypes from Brown et al. (2008) to calculate genome wide associations (Figure A.2 and A.3) and reproduced known sorghum height QTL (Morris et al., 2013; Higgins et al., 2014). As such, the recalibrated reduced representation variants produced by the RIG workflow are useful for common downstream analyses, and these analyses are readily executable due to the GATKs use of standard file formats.